def compute_gaussian_likelihoods(lhs, rhs):
"""Computes sample likelihoods."""
num_lhs_samples = lhs.shape.as_list()[-2] - 2
num_rhs_samples = rhs.shape.as_list()[-2] - 2
lhs_means, lhs_stddevs, lhs_samples = tf.split(
lhs, [1, 1, num_lhs_samples], axis=-2)
rhs_means, rhs_stddevs, rhs_samples = tf.split(
rhs, [1, 1, num_rhs_samples], axis=-2)
rhs_likelihoods = distance_utils.compute_gaussian_likelihoods(
lhs_means,
lhs_stddevs,
rhs_samples,
min_stddev=distance_kernel_kwargs.get(
distance_kernel + '_min_stddev', None),
max_squared_mahalanobis_distance=distance_kernel_kwargs.
get(distance_kernel + '_max_squared_mahalanobis_distance',
None),
smoothing=distance_kernel_kwargs.get(
distance_kernel + '_smoothing', None))
lhs_likelihoods = distance_utils.compute_gaussian_likelihoods(
rhs_means,
rhs_stddevs,
lhs_samples,
l2_distance_computer=l2_distance_computer,
min_stddev=distance_kernel_kwargs.get(
distance_kernel + '_min_stddev', None),
max_squared_mahalanobis_distance=distance_kernel_kwargs.
get(distance_kernel + '_max_squared_mahalanobis_distance',
None),
smoothing=distance_kernel_kwargs.get(
distance_kernel + '_smoothing', None))
return (rhs_likelihoods + lhs_likelihoods) / 2.0
def test_compute_gaussian_likelihoods(self):
# Shape = [2, 1, 1].
means = tf.constant([[[1.0]], [[2.0]]])
# Shape = [2, 1, 1].
stddevs = tf.constant([[[1.0]], [[2.0]]])
# Shape = [2, 3, 1].
samples = tf.constant([[[1.0], [2.0], [3.0]], [[2.0], [4.0], [6.0]]])
# Shape = [2, 3, 1].
likelihoods = distance_utils.compute_gaussian_likelihoods(
means, stddevs, samples)
self.assertAllClose(
likelihoods, [[[1.0], [0.3173], [0.0455]], [[1.0], [0.3173], [0.0455]]],
atol=1e-4)